Phase-Only Pattern Synthesis of Domino Tiling Irregular Arrays Based on Deep Reinforcement Learning and Gradient-descent Optimization

The irregular subarray design technique is gaining increasing attention for its excellent compromise between cost-effective design and good beamforming performance. In this article, a novel framework that incorporates a deep reinforcement learning (DRL) technique with a gradient-descent optimization method for phase-only pattern synthesis of domino-shaped irregular arrays is proposed. By employing the deep Q-network (DQN) technique to tile domino-shaped subarrays into an array aperture, the exact phase distribution over the aperture can be derived. Furthermore, to get better performance of the scanning radiation patterns as well as improve the optimization efficiency, a novel phase-only gradient-descent optimization method is integrated with the DRL training procedure to effectively evaluate the pattern synthesis performance of the tiling configuration. Through the proposed framework by combining DQN and phase-only gradient-descent iterative algorithm, the optimal configuration for the target radiation pattern can be efficiently optimized since the subarray tiling and optimization can be realized simultaneously. The good performance of the proposed framework is demonstrated through several case studies involving 16×16, 32×32, and 64×64-dimensional domino-shaped arrays. Furthermore, full-wave simulations of both reflective metasurface arrays and antenna phased arrays, as well as measurement results from reflective metasurface arrays, are utilized to verify the calculations, thereby demonstrating the practical reliability of the developed method.